2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
29 #include <drm/radeon_drm.h>
31 #include "radeon_trace.h"
35 * GPUVM is similar to the legacy gart on older asics, however
36 * rather than there being a single global gart table
37 * for the entire GPU, there are multiple VM page tables active
38 * at any given time. The VM page tables can contain a mix
39 * vram pages and system memory pages and system memory pages
40 * can be mapped as snooped (cached system pages) or unsnooped
41 * (uncached system pages).
42 * Each VM has an ID associated with it and there is a page table
43 * associated with each VMID. When execting a command buffer,
44 * the kernel tells the the ring what VMID to use for that command
45 * buffer. VMIDs are allocated dynamically as commands are submitted.
46 * The userspace drivers maintain their own address space and the kernel
47 * sets up their pages tables accordingly when they submit their
48 * command buffers and a VMID is assigned.
49 * Cayman/Trinity support up to 8 active VMs at any given time;
54 * radeon_vm_num_pde - return the number of page directory entries
56 * @rdev: radeon_device pointer
58 * Calculate the number of page directory entries (cayman+).
60 static unsigned radeon_vm_num_pdes(struct radeon_device *rdev)
62 return rdev->vm_manager.max_pfn >> radeon_vm_block_size;
66 * radeon_vm_directory_size - returns the size of the page directory in bytes
68 * @rdev: radeon_device pointer
70 * Calculate the size of the page directory in bytes (cayman+).
72 static unsigned radeon_vm_directory_size(struct radeon_device *rdev)
74 return RADEON_GPU_PAGE_ALIGN(radeon_vm_num_pdes(rdev) * 8);
78 * radeon_vm_manager_init - init the vm manager
80 * @rdev: radeon_device pointer
82 * Init the vm manager (cayman+).
83 * Returns 0 for success, error for failure.
85 int radeon_vm_manager_init(struct radeon_device *rdev)
89 if (!rdev->vm_manager.enabled) {
90 r = radeon_asic_vm_init(rdev);
94 rdev->vm_manager.enabled = true;
100 * radeon_vm_manager_fini - tear down the vm manager
102 * @rdev: radeon_device pointer
104 * Tear down the VM manager (cayman+).
106 void radeon_vm_manager_fini(struct radeon_device *rdev)
110 if (!rdev->vm_manager.enabled)
113 for (i = 0; i < RADEON_NUM_VM; ++i)
114 radeon_fence_unref(&rdev->vm_manager.active[i]);
115 radeon_asic_vm_fini(rdev);
116 rdev->vm_manager.enabled = false;
120 * radeon_vm_get_bos - add the vm BOs to a validation list
122 * @vm: vm providing the BOs
123 * @head: head of validation list
125 * Add the page directory to the list of BOs to
126 * validate for command submission (cayman+).
128 struct radeon_cs_reloc *radeon_vm_get_bos(struct radeon_device *rdev,
129 struct radeon_vm *vm,
130 struct list_head *head)
132 struct radeon_cs_reloc *list;
135 list = kmalloc_array(vm->max_pde_used + 2,
136 sizeof(struct radeon_cs_reloc), GFP_KERNEL);
140 /* add the vm page table to the list */
142 list[0].robj = vm->page_directory;
143 list[0].prefered_domains = RADEON_GEM_DOMAIN_VRAM;
144 list[0].allowed_domains = RADEON_GEM_DOMAIN_VRAM;
145 list[0].tv.bo = &vm->page_directory->tbo;
146 list[0].tiling_flags = 0;
148 list_add(&list[0].tv.head, head);
150 for (i = 0, idx = 1; i <= vm->max_pde_used; i++) {
151 if (!vm->page_tables[i].bo)
154 list[idx].gobj = NULL;
155 list[idx].robj = vm->page_tables[i].bo;
156 list[idx].prefered_domains = RADEON_GEM_DOMAIN_VRAM;
157 list[idx].allowed_domains = RADEON_GEM_DOMAIN_VRAM;
158 list[idx].tv.bo = &list[idx].robj->tbo;
159 list[idx].tiling_flags = 0;
160 list[idx].handle = 0;
161 list_add(&list[idx++].tv.head, head);
168 * radeon_vm_grab_id - allocate the next free VMID
170 * @rdev: radeon_device pointer
171 * @vm: vm to allocate id for
172 * @ring: ring we want to submit job to
174 * Allocate an id for the vm (cayman+).
175 * Returns the fence we need to sync to (if any).
177 * Global and local mutex must be locked!
179 struct radeon_fence *radeon_vm_grab_id(struct radeon_device *rdev,
180 struct radeon_vm *vm, int ring)
182 struct radeon_fence *best[RADEON_NUM_RINGS] = {};
183 unsigned choices[2] = {};
186 /* check if the id is still valid */
187 if (vm->last_id_use && vm->last_id_use == rdev->vm_manager.active[vm->id])
190 /* we definately need to flush */
191 radeon_fence_unref(&vm->last_flush);
193 /* skip over VMID 0, since it is the system VM */
194 for (i = 1; i < rdev->vm_manager.nvm; ++i) {
195 struct radeon_fence *fence = rdev->vm_manager.active[i];
198 /* found a free one */
200 trace_radeon_vm_grab_id(vm->id, ring);
204 if (radeon_fence_is_earlier(fence, best[fence->ring])) {
205 best[fence->ring] = fence;
206 choices[fence->ring == ring ? 0 : 1] = i;
210 for (i = 0; i < 2; ++i) {
213 trace_radeon_vm_grab_id(vm->id, ring);
214 return rdev->vm_manager.active[choices[i]];
218 /* should never happen */
224 * radeon_vm_flush - hardware flush the vm
226 * @rdev: radeon_device pointer
227 * @vm: vm we want to flush
228 * @ring: ring to use for flush
230 * Flush the vm (cayman+).
232 * Global and local mutex must be locked!
234 void radeon_vm_flush(struct radeon_device *rdev,
235 struct radeon_vm *vm,
238 uint64_t pd_addr = radeon_bo_gpu_offset(vm->page_directory);
240 /* if we can't remember our last VM flush then flush now! */
241 /* XXX figure out why we have to flush all the time */
242 if (!vm->last_flush || true || pd_addr != vm->pd_gpu_addr) {
243 trace_radeon_vm_flush(pd_addr, ring, vm->id);
244 vm->pd_gpu_addr = pd_addr;
245 radeon_ring_vm_flush(rdev, ring, vm);
250 * radeon_vm_fence - remember fence for vm
252 * @rdev: radeon_device pointer
253 * @vm: vm we want to fence
254 * @fence: fence to remember
256 * Fence the vm (cayman+).
257 * Set the fence used to protect page table and id.
259 * Global and local mutex must be locked!
261 void radeon_vm_fence(struct radeon_device *rdev,
262 struct radeon_vm *vm,
263 struct radeon_fence *fence)
265 radeon_fence_unref(&vm->fence);
266 vm->fence = radeon_fence_ref(fence);
268 radeon_fence_unref(&rdev->vm_manager.active[vm->id]);
269 rdev->vm_manager.active[vm->id] = radeon_fence_ref(fence);
271 radeon_fence_unref(&vm->last_id_use);
272 vm->last_id_use = radeon_fence_ref(fence);
274 /* we just flushed the VM, remember that */
276 vm->last_flush = radeon_fence_ref(fence);
280 * radeon_vm_bo_find - find the bo_va for a specific vm & bo
283 * @bo: requested buffer object
285 * Find @bo inside the requested vm (cayman+).
286 * Search inside the @bos vm list for the requested vm
287 * Returns the found bo_va or NULL if none is found
289 * Object has to be reserved!
291 struct radeon_bo_va *radeon_vm_bo_find(struct radeon_vm *vm,
292 struct radeon_bo *bo)
294 struct radeon_bo_va *bo_va;
296 list_for_each_entry(bo_va, &bo->va, bo_list) {
297 if (bo_va->vm == vm) {
305 * radeon_vm_bo_add - add a bo to a specific vm
307 * @rdev: radeon_device pointer
309 * @bo: radeon buffer object
311 * Add @bo into the requested vm (cayman+).
312 * Add @bo to the list of bos associated with the vm
313 * Returns newly added bo_va or NULL for failure
315 * Object has to be reserved!
317 struct radeon_bo_va *radeon_vm_bo_add(struct radeon_device *rdev,
318 struct radeon_vm *vm,
319 struct radeon_bo *bo)
321 struct radeon_bo_va *bo_va;
323 bo_va = kzalloc(sizeof(struct radeon_bo_va), GFP_KERNEL);
332 bo_va->valid = false;
333 bo_va->ref_count = 1;
334 INIT_LIST_HEAD(&bo_va->bo_list);
335 INIT_LIST_HEAD(&bo_va->vm_list);
336 INIT_LIST_HEAD(&bo_va->vm_status);
338 mutex_lock(&vm->mutex);
339 list_add(&bo_va->vm_list, &vm->va);
340 list_add_tail(&bo_va->bo_list, &bo->va);
341 mutex_unlock(&vm->mutex);
347 * radeon_vm_clear_bo - initially clear the page dir/table
349 * @rdev: radeon_device pointer
352 static int radeon_vm_clear_bo(struct radeon_device *rdev,
353 struct radeon_bo *bo)
355 struct ttm_validate_buffer tv;
356 struct ww_acquire_ctx ticket;
357 struct list_head head;
363 memset(&tv, 0, sizeof(tv));
366 INIT_LIST_HEAD(&head);
367 list_add(&tv.head, &head);
369 r = ttm_eu_reserve_buffers(&ticket, &head);
373 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
377 addr = radeon_bo_gpu_offset(bo);
378 entries = radeon_bo_size(bo) / 8;
380 r = radeon_ib_get(rdev, R600_RING_TYPE_DMA_INDEX, &ib,
381 NULL, entries * 2 + 64);
387 radeon_asic_vm_set_page(rdev, &ib, addr, 0, entries, 0, 0);
389 r = radeon_ib_schedule(rdev, &ib, NULL);
393 ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
394 radeon_ib_free(rdev, &ib);
399 ttm_eu_backoff_reservation(&ticket, &head);
404 * radeon_vm_bo_set_addr - set bos virtual address inside a vm
406 * @rdev: radeon_device pointer
407 * @bo_va: bo_va to store the address
408 * @soffset: requested offset of the buffer in the VM address space
409 * @flags: attributes of pages (read/write/valid/etc.)
411 * Set offset of @bo_va (cayman+).
412 * Validate and set the offset requested within the vm address space.
413 * Returns 0 for success, error for failure.
415 * Object has to be reserved!
417 int radeon_vm_bo_set_addr(struct radeon_device *rdev,
418 struct radeon_bo_va *bo_va,
422 uint64_t size = radeon_bo_size(bo_va->bo);
423 uint64_t eoffset, last_offset = 0;
424 struct radeon_vm *vm = bo_va->vm;
425 struct radeon_bo_va *tmp;
426 struct list_head *head;
427 unsigned last_pfn, pt_idx;
431 /* make sure object fit at this offset */
432 eoffset = soffset + size;
433 if (soffset >= eoffset) {
437 last_pfn = eoffset / RADEON_GPU_PAGE_SIZE;
438 if (last_pfn > rdev->vm_manager.max_pfn) {
439 dev_err(rdev->dev, "va above limit (0x%08X > 0x%08X)\n",
440 last_pfn, rdev->vm_manager.max_pfn);
445 eoffset = last_pfn = 0;
448 mutex_lock(&vm->mutex);
451 list_for_each_entry(tmp, &vm->va, vm_list) {
453 /* skip over currently modified bo */
457 if (soffset >= last_offset && eoffset <= tmp->soffset) {
458 /* bo can be added before this one */
461 if (eoffset > tmp->soffset && soffset < tmp->eoffset) {
462 /* bo and tmp overlap, invalid offset */
463 dev_err(rdev->dev, "bo %p va 0x%08X conflict with (bo %p 0x%08X 0x%08X)\n",
464 bo_va->bo, (unsigned)bo_va->soffset, tmp->bo,
465 (unsigned)tmp->soffset, (unsigned)tmp->eoffset);
466 mutex_unlock(&vm->mutex);
469 last_offset = tmp->eoffset;
470 head = &tmp->vm_list;
473 if (bo_va->soffset) {
474 /* add a clone of the bo_va to clear the old address */
475 tmp = kzalloc(sizeof(struct radeon_bo_va), GFP_KERNEL);
477 mutex_unlock(&vm->mutex);
480 tmp->soffset = bo_va->soffset;
481 tmp->eoffset = bo_va->eoffset;
483 list_add(&tmp->vm_status, &vm->freed);
486 bo_va->soffset = soffset;
487 bo_va->eoffset = eoffset;
488 bo_va->flags = flags;
489 bo_va->valid = false;
490 list_move(&bo_va->vm_list, head);
492 soffset = (soffset / RADEON_GPU_PAGE_SIZE) >> radeon_vm_block_size;
493 eoffset = (eoffset / RADEON_GPU_PAGE_SIZE) >> radeon_vm_block_size;
495 BUG_ON(eoffset >= radeon_vm_num_pdes(rdev));
497 if (eoffset > vm->max_pde_used)
498 vm->max_pde_used = eoffset;
500 radeon_bo_unreserve(bo_va->bo);
502 /* walk over the address space and allocate the page tables */
503 for (pt_idx = soffset; pt_idx <= eoffset; ++pt_idx) {
504 struct radeon_bo *pt;
506 if (vm->page_tables[pt_idx].bo)
509 /* drop mutex to allocate and clear page table */
510 mutex_unlock(&vm->mutex);
512 r = radeon_bo_create(rdev, RADEON_VM_PTE_COUNT * 8,
513 RADEON_GPU_PAGE_SIZE, true,
514 RADEON_GEM_DOMAIN_VRAM, 0, NULL, &pt);
518 r = radeon_vm_clear_bo(rdev, pt);
520 radeon_bo_unref(&pt);
521 radeon_bo_reserve(bo_va->bo, false);
525 /* aquire mutex again */
526 mutex_lock(&vm->mutex);
527 if (vm->page_tables[pt_idx].bo) {
528 /* someone else allocated the pt in the meantime */
529 mutex_unlock(&vm->mutex);
530 radeon_bo_unref(&pt);
531 mutex_lock(&vm->mutex);
535 vm->page_tables[pt_idx].addr = 0;
536 vm->page_tables[pt_idx].bo = pt;
539 mutex_unlock(&vm->mutex);
540 return radeon_bo_reserve(bo_va->bo, false);
544 * radeon_vm_map_gart - get the physical address of a gart page
546 * @rdev: radeon_device pointer
547 * @addr: the unmapped addr
549 * Look up the physical address of the page that the pte resolves
551 * Returns the physical address of the page.
553 uint64_t radeon_vm_map_gart(struct radeon_device *rdev, uint64_t addr)
557 /* page table offset */
558 result = rdev->gart.pages_addr[addr >> PAGE_SHIFT];
560 /* in case cpu page size != gpu page size*/
561 result |= addr & (~PAGE_MASK);
567 * radeon_vm_page_flags - translate page flags to what the hw uses
569 * @flags: flags comming from userspace
571 * Translate the flags the userspace ABI uses to hw flags.
573 static uint32_t radeon_vm_page_flags(uint32_t flags)
575 uint32_t hw_flags = 0;
576 hw_flags |= (flags & RADEON_VM_PAGE_VALID) ? R600_PTE_VALID : 0;
577 hw_flags |= (flags & RADEON_VM_PAGE_READABLE) ? R600_PTE_READABLE : 0;
578 hw_flags |= (flags & RADEON_VM_PAGE_WRITEABLE) ? R600_PTE_WRITEABLE : 0;
579 if (flags & RADEON_VM_PAGE_SYSTEM) {
580 hw_flags |= R600_PTE_SYSTEM;
581 hw_flags |= (flags & RADEON_VM_PAGE_SNOOPED) ? R600_PTE_SNOOPED : 0;
587 * radeon_vm_update_pdes - make sure that page directory is valid
589 * @rdev: radeon_device pointer
591 * @start: start of GPU address range
592 * @end: end of GPU address range
594 * Allocates new page tables if necessary
595 * and updates the page directory (cayman+).
596 * Returns 0 for success, error for failure.
598 * Global and local mutex must be locked!
600 int radeon_vm_update_page_directory(struct radeon_device *rdev,
601 struct radeon_vm *vm)
603 struct radeon_bo *pd = vm->page_directory;
604 uint64_t pd_addr = radeon_bo_gpu_offset(pd);
605 uint32_t incr = RADEON_VM_PTE_COUNT * 8;
606 uint64_t last_pde = ~0, last_pt = ~0;
607 unsigned count = 0, pt_idx, ndw;
614 /* assume the worst case */
615 ndw += vm->max_pde_used * 16;
617 /* update too big for an IB */
621 r = radeon_ib_get(rdev, R600_RING_TYPE_DMA_INDEX, &ib, NULL, ndw * 4);
626 /* walk over the address space and update the page directory */
627 for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) {
628 struct radeon_bo *bo = vm->page_tables[pt_idx].bo;
634 pt = radeon_bo_gpu_offset(bo);
635 if (vm->page_tables[pt_idx].addr == pt)
637 vm->page_tables[pt_idx].addr = pt;
639 pde = pd_addr + pt_idx * 8;
640 if (((last_pde + 8 * count) != pde) ||
641 ((last_pt + incr * count) != pt)) {
644 radeon_asic_vm_set_page(rdev, &ib, last_pde,
645 last_pt, count, incr,
658 radeon_asic_vm_set_page(rdev, &ib, last_pde, last_pt, count,
659 incr, R600_PTE_VALID);
661 if (ib.length_dw != 0) {
662 radeon_semaphore_sync_to(ib.semaphore, pd->tbo.sync_obj);
663 radeon_semaphore_sync_to(ib.semaphore, vm->last_id_use);
664 r = radeon_ib_schedule(rdev, &ib, NULL);
666 radeon_ib_free(rdev, &ib);
669 radeon_fence_unref(&vm->fence);
670 vm->fence = radeon_fence_ref(ib.fence);
671 radeon_fence_unref(&vm->last_flush);
673 radeon_ib_free(rdev, &ib);
679 * radeon_vm_frag_ptes - add fragment information to PTEs
681 * @rdev: radeon_device pointer
682 * @ib: IB for the update
683 * @pe_start: first PTE to handle
684 * @pe_end: last PTE to handle
685 * @addr: addr those PTEs should point to
686 * @flags: hw mapping flags
688 * Global and local mutex must be locked!
690 static void radeon_vm_frag_ptes(struct radeon_device *rdev,
691 struct radeon_ib *ib,
692 uint64_t pe_start, uint64_t pe_end,
693 uint64_t addr, uint32_t flags)
696 * The MC L1 TLB supports variable sized pages, based on a fragment
697 * field in the PTE. When this field is set to a non-zero value, page
698 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
699 * flags are considered valid for all PTEs within the fragment range
700 * and corresponding mappings are assumed to be physically contiguous.
702 * The L1 TLB can store a single PTE for the whole fragment,
703 * significantly increasing the space available for translation
704 * caching. This leads to large improvements in throughput when the
705 * TLB is under pressure.
707 * The L2 TLB distributes small and large fragments into two
708 * asymmetric partitions. The large fragment cache is significantly
709 * larger. Thus, we try to use large fragments wherever possible.
710 * Userspace can support this by aligning virtual base address and
711 * allocation size to the fragment size.
714 /* NI is optimized for 256KB fragments, SI and newer for 64KB */
715 uint64_t frag_flags = rdev->family == CHIP_CAYMAN ?
716 R600_PTE_FRAG_256KB : R600_PTE_FRAG_64KB;
717 uint64_t frag_align = rdev->family == CHIP_CAYMAN ? 0x200 : 0x80;
719 uint64_t frag_start = ALIGN(pe_start, frag_align);
720 uint64_t frag_end = pe_end & ~(frag_align - 1);
724 /* system pages are non continuously */
725 if ((flags & R600_PTE_SYSTEM) || !(flags & R600_PTE_VALID) ||
726 (frag_start >= frag_end)) {
728 count = (pe_end - pe_start) / 8;
729 radeon_asic_vm_set_page(rdev, ib, pe_start, addr, count,
730 RADEON_GPU_PAGE_SIZE, flags);
734 /* handle the 4K area at the beginning */
735 if (pe_start != frag_start) {
736 count = (frag_start - pe_start) / 8;
737 radeon_asic_vm_set_page(rdev, ib, pe_start, addr, count,
738 RADEON_GPU_PAGE_SIZE, flags);
739 addr += RADEON_GPU_PAGE_SIZE * count;
742 /* handle the area in the middle */
743 count = (frag_end - frag_start) / 8;
744 radeon_asic_vm_set_page(rdev, ib, frag_start, addr, count,
745 RADEON_GPU_PAGE_SIZE, flags | frag_flags);
747 /* handle the 4K area at the end */
748 if (frag_end != pe_end) {
749 addr += RADEON_GPU_PAGE_SIZE * count;
750 count = (pe_end - frag_end) / 8;
751 radeon_asic_vm_set_page(rdev, ib, frag_end, addr, count,
752 RADEON_GPU_PAGE_SIZE, flags);
757 * radeon_vm_update_ptes - make sure that page tables are valid
759 * @rdev: radeon_device pointer
761 * @start: start of GPU address range
762 * @end: end of GPU address range
763 * @dst: destination address to map to
764 * @flags: mapping flags
766 * Update the page tables in the range @start - @end (cayman+).
768 * Global and local mutex must be locked!
770 static void radeon_vm_update_ptes(struct radeon_device *rdev,
771 struct radeon_vm *vm,
772 struct radeon_ib *ib,
773 uint64_t start, uint64_t end,
774 uint64_t dst, uint32_t flags)
776 uint64_t mask = RADEON_VM_PTE_COUNT - 1;
777 uint64_t last_pte = ~0, last_dst = ~0;
781 start = start / RADEON_GPU_PAGE_SIZE;
782 end = end / RADEON_GPU_PAGE_SIZE;
784 /* walk over the address space and update the page tables */
785 for (addr = start; addr < end; ) {
786 uint64_t pt_idx = addr >> radeon_vm_block_size;
787 struct radeon_bo *pt = vm->page_tables[pt_idx].bo;
791 radeon_semaphore_sync_to(ib->semaphore, pt->tbo.sync_obj);
793 if ((addr & ~mask) == (end & ~mask))
796 nptes = RADEON_VM_PTE_COUNT - (addr & mask);
798 pte = radeon_bo_gpu_offset(pt);
799 pte += (addr & mask) * 8;
801 if ((last_pte + 8 * count) != pte) {
804 radeon_vm_frag_ptes(rdev, ib, last_pte,
805 last_pte + 8 * count,
817 dst += nptes * RADEON_GPU_PAGE_SIZE;
821 radeon_vm_frag_ptes(rdev, ib, last_pte,
822 last_pte + 8 * count,
828 * radeon_vm_bo_update - map a bo into the vm page table
830 * @rdev: radeon_device pointer
832 * @bo: radeon buffer object
835 * Fill in the page table entries for @bo (cayman+).
836 * Returns 0 for success, -EINVAL for failure.
838 * Object have to be reserved and mutex must be locked!
840 int radeon_vm_bo_update(struct radeon_device *rdev,
841 struct radeon_bo_va *bo_va,
842 struct ttm_mem_reg *mem)
844 struct radeon_vm *vm = bo_va->vm;
851 if (!bo_va->soffset) {
852 dev_err(rdev->dev, "bo %p don't has a mapping in vm %p\n",
857 if ((bo_va->valid && mem) || (!bo_va->valid && mem == NULL))
860 bo_va->flags &= ~RADEON_VM_PAGE_VALID;
861 bo_va->flags &= ~RADEON_VM_PAGE_SYSTEM;
862 bo_va->flags &= ~RADEON_VM_PAGE_SNOOPED;
864 addr = mem->start << PAGE_SHIFT;
865 if (mem->mem_type != TTM_PL_SYSTEM) {
866 bo_va->flags |= RADEON_VM_PAGE_VALID;
869 if (mem->mem_type == TTM_PL_TT) {
870 bo_va->flags |= RADEON_VM_PAGE_SYSTEM;
871 if (!(bo_va->bo->flags & (RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC)))
872 bo_va->flags |= RADEON_VM_PAGE_SNOOPED;
875 addr += rdev->vm_manager.vram_base_offset;
879 bo_va->valid = false;
882 trace_radeon_vm_bo_update(bo_va);
884 nptes = (bo_va->eoffset - bo_va->soffset) / RADEON_GPU_PAGE_SIZE;
889 if (radeon_vm_block_size > 11)
890 /* reserve space for one header for every 2k dwords */
891 ndw += (nptes >> 11) * 4;
893 /* reserve space for one header for
894 every (1 << BLOCK_SIZE) entries */
895 ndw += (nptes >> radeon_vm_block_size) * 4;
897 /* reserve space for pte addresses */
900 /* update too big for an IB */
904 r = radeon_ib_get(rdev, R600_RING_TYPE_DMA_INDEX, &ib, NULL, ndw * 4);
909 radeon_vm_update_ptes(rdev, vm, &ib, bo_va->soffset, bo_va->eoffset,
910 addr, radeon_vm_page_flags(bo_va->flags));
912 radeon_semaphore_sync_to(ib.semaphore, vm->fence);
913 r = radeon_ib_schedule(rdev, &ib, NULL);
915 radeon_ib_free(rdev, &ib);
918 radeon_fence_unref(&vm->fence);
919 vm->fence = radeon_fence_ref(ib.fence);
920 radeon_ib_free(rdev, &ib);
921 radeon_fence_unref(&vm->last_flush);
927 * radeon_vm_clear_freed - clear freed BOs in the PT
929 * @rdev: radeon_device pointer
932 * Make sure all freed BOs are cleared in the PT.
933 * Returns 0 for success.
935 * PTs have to be reserved and mutex must be locked!
937 int radeon_vm_clear_freed(struct radeon_device *rdev,
938 struct radeon_vm *vm)
940 struct radeon_bo_va *bo_va, *tmp;
943 list_for_each_entry_safe(bo_va, tmp, &vm->freed, vm_status) {
944 list_del(&bo_va->vm_status);
945 r = radeon_vm_bo_update(rdev, bo_va, NULL);
955 * radeon_vm_bo_rmv - remove a bo to a specific vm
957 * @rdev: radeon_device pointer
958 * @bo_va: requested bo_va
960 * Remove @bo_va->bo from the requested vm (cayman+).
962 * Object have to be reserved!
964 void radeon_vm_bo_rmv(struct radeon_device *rdev,
965 struct radeon_bo_va *bo_va)
967 struct radeon_vm *vm = bo_va->vm;
969 list_del(&bo_va->bo_list);
971 mutex_lock(&vm->mutex);
972 list_del(&bo_va->vm_list);
974 if (bo_va->soffset) {
976 list_add(&bo_va->vm_status, &vm->freed);
981 mutex_unlock(&vm->mutex);
985 * radeon_vm_bo_invalidate - mark the bo as invalid
987 * @rdev: radeon_device pointer
989 * @bo: radeon buffer object
991 * Mark @bo as invalid (cayman+).
993 void radeon_vm_bo_invalidate(struct radeon_device *rdev,
994 struct radeon_bo *bo)
996 struct radeon_bo_va *bo_va;
998 list_for_each_entry(bo_va, &bo->va, bo_list) {
999 bo_va->valid = false;
1004 * radeon_vm_init - initialize a vm instance
1006 * @rdev: radeon_device pointer
1009 * Init @vm fields (cayman+).
1011 int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm)
1013 const unsigned align = min(RADEON_VM_PTB_ALIGN_SIZE,
1014 RADEON_VM_PTE_COUNT * 8);
1015 unsigned pd_size, pd_entries, pts_size;
1019 vm->ib_bo_va = NULL;
1021 vm->last_flush = NULL;
1022 vm->last_id_use = NULL;
1023 mutex_init(&vm->mutex);
1024 INIT_LIST_HEAD(&vm->va);
1025 INIT_LIST_HEAD(&vm->freed);
1027 pd_size = radeon_vm_directory_size(rdev);
1028 pd_entries = radeon_vm_num_pdes(rdev);
1030 /* allocate page table array */
1031 pts_size = pd_entries * sizeof(struct radeon_vm_pt);
1032 vm->page_tables = kzalloc(pts_size, GFP_KERNEL);
1033 if (vm->page_tables == NULL) {
1034 DRM_ERROR("Cannot allocate memory for page table array\n");
1038 r = radeon_bo_create(rdev, pd_size, align, true,
1039 RADEON_GEM_DOMAIN_VRAM, 0, NULL,
1040 &vm->page_directory);
1044 r = radeon_vm_clear_bo(rdev, vm->page_directory);
1046 radeon_bo_unref(&vm->page_directory);
1047 vm->page_directory = NULL;
1055 * radeon_vm_fini - tear down a vm instance
1057 * @rdev: radeon_device pointer
1060 * Tear down @vm (cayman+).
1061 * Unbind the VM and remove all bos from the vm bo list
1063 void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
1065 struct radeon_bo_va *bo_va, *tmp;
1068 if (!list_empty(&vm->va)) {
1069 dev_err(rdev->dev, "still active bo inside vm\n");
1071 list_for_each_entry_safe(bo_va, tmp, &vm->va, vm_list) {
1072 list_del_init(&bo_va->vm_list);
1073 r = radeon_bo_reserve(bo_va->bo, false);
1075 list_del_init(&bo_va->bo_list);
1076 radeon_bo_unreserve(bo_va->bo);
1080 list_for_each_entry_safe(bo_va, tmp, &vm->freed, vm_status)
1083 for (i = 0; i < radeon_vm_num_pdes(rdev); i++)
1084 radeon_bo_unref(&vm->page_tables[i].bo);
1085 kfree(vm->page_tables);
1087 radeon_bo_unref(&vm->page_directory);
1089 radeon_fence_unref(&vm->fence);
1090 radeon_fence_unref(&vm->last_flush);
1091 radeon_fence_unref(&vm->last_id_use);
1093 mutex_destroy(&vm->mutex);